Here is every byte that leaves your machine, for each tool
"Private" is a word every dictation app uses and none of them defines. This page defines it as a table: for each kind of thing a dictation tool could send off your machine, what is sent, where it goes, and when. Deixis's table is short because speech recognition happens inside the app on your own device, on Windows, Mac, iPhone and Android alike, and there is no account: audio never leaves, and with no cleanup endpoint configured nothing but the update check does.
The other tools' tables are filled in from their own documentation and nothing else. Where a document says nothing about a row, the row says so and links to the document that was read, with the date. That is a claim about their documentation, not about their software, and it is the only kind of claim we can make about code we cannot see.
Deixis, in full
Everything that leaves a Deixis install, and when
The README's own table, row for row. The bottom four rows are off until you configure a destination; the update check is the one call a fresh install makes on its own. Nothing in this table carries audio.
| What | What is sent | Where | When |
|---|---|---|---|
| Update check | nothing but the request (the version comparison happens on your machine) | api.deixis-voice.com/release/ | every launch, and on demand in Settings → Updates |
| Speech model download | nothing but the request | huggingface.co | first run only, when you press Download |
| Cleanup | the transcript, your vocabulary terms, and where it is going: the focused app, the control and field name, and your previous utterance in that field. Not the window title | the endpoint you configure | only with an endpoint set, and by default only when rules are not enough |
| Translate | the selected text, and your vocabulary terms | the same endpoint, /translate | when you press the translate key |
| Meeting notes | the meeting transcript | the same endpoint, /summarize | at the end of a meeting, only with an endpoint set |
| Integrations and webhook buttons | the finished note or translation, and for a grab its path and filename — never the image itself | the URL you enter | on finish, only for destinations you add |
Never sent, under any setting
Audio. Screen grabs. Recorded meeting WAVs. Window titles. There is no setting that turns those on, because there is no code that reads them for the network.
Off until you configure them
With no endpoint set, dictation falls back to the local rules engine and Deixis works entirely offline. Integrations are your own URLs and only exist once you add one.
The other tools
The same rows, from each vendor's own documentation
Eight rows per tool. A row reads "not documented" when the vendor's privacy policy or documentation did not address it on the date shown; that is a fact about the document, and the link is to the document.
Aqua Voice (macOS, Windows, iOS)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Your voice, to be turned into text: the Windows page says Aqua processes audio in the cloud, where its Avalon model runs, with no offline mode, and the FAQ says audio is processed in the cloud and never sold or shared for advertising. Realtime Mode streams audio to a third-party speech provider. Its history guide says recent sessions' audio is also kept on your device and deleted after three days. | Aqua's cloud, and a third-party provider in Realtime Mode · every dictation | checked 2026-09-05 |
| The dictated text | The transcript is produced in the cloud. By default, the FAQ says, transcripts may be retained to improve the model; with Privacy Mode on they are not, and the home page says nothing is then stored on Aqua's servers. | Aqua's servers · every dictation | checked 2026-09-05 |
| Context around the text | Deep Context, its screen reading, stays off until you enable it and, the FAQ says, is not stored; when it is on, the FAQ says it reads what is on screen so that code, names and jargon come out correctly cased. The pages do not say what form the screen content takes when it is sent. | not stated · only with Deep Context on | checked 2026-09-05 |
| Telemetry and analytics | Session metadata (timestamps, device type, performance metrics) and technical data (IP address, browser type, operating system). For advertising measurement and app-install attribution, a first-party identifier, and a hashed version of your email address sent with the event to the advertising platforms the policy names. | Aqua, its attribution provider and the named advertising platforms · not stated | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | not documented | checked 2026-09-05 | |
| Account and sign-in | Name, email address, billing address and payment information | Aqua · when you create the account and when you pay | checked 2026-09-05 |
| Integrations | not documented | checked 2026-09-05 | |
Facts about Aqua Voice checked on 2026-09-05 against aquavoice.com.
Dragon Professional (Windows 10 and 11, Windows Server 2016, 2019 and 2022)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Nothing for recognition: v16 is the 'locally installed' Dragon, and the 15.6 release notes say 'audio data collection has been discontinued'. Nuance's general privacy statement, which is not specific to Dragon, still says it 'may capture your voice and the words that you speak' when you use its voice recognition technology, and that 'snippets of your voice data may be manually reviewed'. | Nuance, where it applies · not stated for this product | checked 2026-09-05 |
| The dictated text | Up to 500 MB of 'data and text from your dictation sessions', only if you opt in to Data Collection; 'No personal information is ever sent to Nuance' | Nuance · only if you opted in, during profile setup or later in Options | checked 2026-09-05 |
| Context around the text | not documented | checked 2026-09-05 | |
| Telemetry and analytics | Google Analytics metrics 'on how you use Dragon', if you switch on User Experience Collection; 'No personal information is ever collected, and you can stop participating at any time' | Google Analytics, for Nuance · only if you opted in | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | not documented | checked 2026-09-05 | |
| Account and sign-in | A product activation, 'a quick anonymous process' that needs an internet connection; managed deployments also connect to the Nuance Management Center | Nuance · at activation, and for managed deployments whenever the client connects | checked 2026-09-05 |
| Integrations | User customisations such as auto-texts, custom words and commands, synchronised between Dragon clients on PC, iOS and Android | the Nuance Management Center · in enterprise deployments that use the Management Center | checked 2026-09-05 |
Facts about Dragon Professional checked on 2026-09-05 against dragon.nuance.com.
MacWhisper (macOS)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Nothing, by default: the privacy policy says no data, audio, text or other, leaves your device. Audio goes to a cloud transcription provider only if you connect one under your own key. | a provider you choose, if any · only when you connect one | checked 2026-09-05 |
| The dictated text | Nothing, by default. With a language-model provider connected under your own key, the transcript goes to it for the AI features. | a provider you choose, if any · only when you use an AI feature with a cloud provider | checked 2026-09-05 |
| Context around the text | not documented | checked 2026-09-05 | |
| Telemetry and analytics | Anonymous licensing-related analytics, to verify a valid licence | the vendor · not stated | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | not documented | checked 2026-09-05 | |
| Account and sign-in | not documented | checked 2026-09-05 | |
| Integrations | not documented | checked 2026-09-05 | |
Facts about MacWhisper checked on 2026-09-05 against macwhisper.com.
Otter.ai (Web browser, Windows and macOS (12.3 or later) desktop app, iOS 15 or later, Android 5.0 or later, Chrome extension)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Your recordings: what the Notetaker hears in the call, or what the desktop app captures from your microphone and speakers. The desktop app says it 'will automatically start uploading and transcribing the file once a connection is detected'. | Otter's servers on Amazon Web Services in the United States · every recording | checked 2026-09-05 |
| The dictated text | Transcripts, summaries, chat and anything you upload; de-identified recordings and transcriptions 'which may contain Personal Information' may be used to train Otter's AI | Otter, and its AI and data-labelling service providers · always; summaries and AI chat run in Otter's cloud | checked 2026-09-05 |
| Context around the text | Calendar events and contacts when you connect Google or Microsoft, and in virtual meetings the Notetaker 'may take automatic screenshots' of the meeting it is sitting in, not of your desktop | Otter · with a calendar connected, and during meetings the bot records | checked 2026-09-05 |
| Telemetry and analytics | Usage and device information: your IP address, device identifiers and 'web beacons' | Otter, its analytics providers and advertising partners · while you use the service | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | not documented | checked 2026-09-05 | |
| Account and sign-in | Your name, email, phone number and a password | Otter · at sign-up; the desktop app asks you to sign in before it records | checked 2026-09-05 |
| Integrations | Real-time updates to Slack, files you drop in an Otter folder on Dropbox, and calendar access for auto-join | the service you connect · only once you connect one | checked 2026-09-05 |
Facts about Otter.ai checked on 2026-09-05 against otter.ai.
Superwhisper (macOS, Windows, iOS)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Nothing with a local voice model: the local Whisper models run on the device and the privacy policy says no audio is collected. With a cloud voice model — Superwhisper's own servers or Deepgram, all Pro — the audio goes out for transcription; its security docs say those providers keep nothing and do not train on it. | Superwhisper's servers or Deepgram · only with a cloud voice model selected | checked 2026-09-05 |
| The dictated text | The transcript, whenever an AI mode runs: every language model its docs list is a cloud model, from Anthropic, OpenAI, Groq or Superwhisper's own servers, all of them Pro, and its security docs say a local language model is a macOS-only option. | the provider of the language model you picked · on every dictation in an AI mode | checked 2026-09-05 |
| Context around the text | The active window's title and name, text from the active input field, selected text and recently copied text, plus the date, time, your full name and computer name. History shows the full prompt that went, so you can check. | the provider of the language model you picked · in Super Mode, and in any mode with a context source on | checked 2026-09-05 |
| Telemetry and analytics | None, says the privacy policy: no usage data is collected. An enterprise dashboard, off by default until an owner turns it on, counts dictations, mode, model, duration, app version and platform — never audio or text. | the vendor, enterprise plans only · only when an enterprise owner switches it on | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | A model download when you pick one: the free local Whisper models are 75 MB to 500 MB, the Pro ones up to 3 GB. The docs do not say what the request carries. | not stated · when you select a model | checked 2026-09-05 |
| Account and sign-in | A licence key for Pro, handed to you at purchase and pasted into the app. Server-side the vendor keeps an obfuscated email address and anonymised billing records, and says it stores no transcripts, audio, modes or vocabulary. | the vendor and its payment processor · on purchase and activation | checked 2026-09-05 |
| Integrations | Your modes and settings, if you turn on FileSync, through Superwhisper's infrastructure; its docs say this never includes transcription content or audio. | Superwhisper's servers · only with FileSync on | checked 2026-09-05 |
Facts about Superwhisper checked on 2026-09-05 against superwhisper.com.
Talon (Windows 8 or newer, macOS 10.13 or newer, Linux (X11 only))
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Nothing: the licence agreement says its metrics 'do NOT record actual input, speech, text, or screen contents', and the recognition engine is one you install into the app from its Speech Recognition menu | — · — | checked 2026-09-05 |
| The dictated text | Nothing: the same sentence excludes 'input' and 'text' | — · — | checked 2026-09-05 |
| Context around the text | Nothing: the same sentence excludes 'screen contents' | — · — | checked 2026-09-05 |
| Telemetry and analytics | Usage-based metrics, 'mostly numerical Talon app/script performance information and percentage ratios', that 'may be uploaded to a Talon-owned server', and crash reports, which are uploaded automatically on the next start after a crash | Talon's own server · when metrics are uploaded is not stated; a crash report the next time Talon starts | checked 2026-09-05 |
| Update check | Your IP address and app version, 'if you opt into automatic updates' | the Talon update server · periodically while Talon is running, only if you opted in | checked 2026-09-05 |
| Model download | A download of the Conformer speech model from the Speech Recognition menu | not stated · when you install the engine | checked 2026-09-05 |
| Account and sign-in | not documented | checked 2026-09-05 | |
| Integrations | The community packages system 'may log, store, and transmit information about which plugins are installed and their frequency of use' | not stated · when you use community packages | checked 2026-09-05 |
Facts about Talon checked on 2026-09-05 against talonvoice.com.
Typeless (Windows, macOS (Apple Silicon and Intel), iOS, Android)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Your voice. The privacy policy: audio inputs and contextual information are processed in real time on Typeless's cloud servers and immediately discarded once the transcription result is returned to your device. | Typeless's cloud servers · every dictation | checked 2026-09-05 |
| The dictated text | The transcript comes back to your device. History stays on the device unless you enable History cloud sync, in which case history text is stored until you turn it off or delete it. Text handled by third-party LLM providers is configured for zero retention with them, per the policy. | your device; Typeless's servers with cloud sync on; third-party LLM providers for the AI features · with each dictation; cloud sync only when you enable it | checked 2026-09-05 |
| Context around the text | Limited contextual information: the application you are using and relevant text within it, sent with the audio and discarded once the result is returned. The policy says no screen context data is stored on its servers. | Typeless's cloud servers · with each dictation | checked 2026-09-05 |
| Telemetry and analytics | Usage data: IP address, device type, browser type and version, pages visited and system logs; Google Analytics is named as a third-party analytics service. | Typeless and Google Analytics · not stated per feature | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | Nothing to download: recognition runs on Typeless's cloud servers, not on a model on your PC. | — · — | checked 2026-09-05 |
| Account and sign-in | An account is required: the setup guide has you sign in with Google, Apple or your email address. The policy lists email address and, for paid plans, payment details handled by Stripe, Apple or Google. | Typeless · at sign-up, before the first dictation | checked 2026-09-05 |
| Integrations | not documented | checked 2026-09-05 | |
Facts about Typeless checked on 2026-09-05 against typeless.com.
VoiceInk (macOS)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Nothing, by default: the privacy policy says VoiceInk does not collect or transmit any personal data by default and that audio stays on your Mac. Audio goes to a cloud transcription provider only if you add your own key for one and pick its model. | a provider you choose, if any · only when a cloud transcription model is selected | checked 2026-09-05 |
| The dictated text | Nothing, by default. With a cloud enhancement provider connected under your own key, the transcribed text goes to it; the policy says only the text, never the audio, is sent for enhancement. | a provider you choose, if any · only when a mode has AI enhancement on with a cloud provider | checked 2026-09-05 |
| Context around the text | Selected text, the clipboard, and text read off the active window on the Mac — the docs say the screenshot itself is never the payload, only the text extracted from it. Each source is a per-mode switch you turn on yourself, and a plain dictation mode never sends context. | the enhancement provider you chose · only in a mode with that source switched on, when AI enhancement runs | checked 2026-09-05 |
| Telemetry and analytics | not documented | checked 2026-09-05 | |
| Update check | An update check in the background while Auto-check Updates is on, and on demand with Check for Updates. The docs do not say what the request carries. | not stated · in the background, when the setting is on | checked 2026-09-05 |
| Model download | A model download when you press Download on a Whisper or Parakeet model card; the Whisper and Parakeet models are fetched on demand, and Apple Speech is built into macOS 26. The docs do not say what the request carries. | not stated · when you download a model | checked 2026-09-05 |
| Account and sign-in | A licence key, emailed after purchase and pasted into the app; keys and the Macs they cover are managed on the payment provider's portal. The docs do not describe what activation sends. | not stated · when you activate a purchase | checked 2026-09-05 |
| Integrations | not documented | checked 2026-09-05 | |
Facts about VoiceInk checked on 2026-09-05 against tryvoiceink.com.
Whisperstream (Windows 10 or 11, 64-bit)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Nothing: 'Microphone audio is held in random-access memory while each push-to-talk capture is transcribed on your CPU, and is never uploaded to any server' | — · — | checked 2026-09-05 |
| The dictated text | Nothing by default. With AI enhancement on and a cloud provider chosen, 'transcribed text longer than three words is transmitted only to the specific third-party provider you selected' | the provider you chose, under your own key · only once you enable enhancement with a cloud provider | checked 2026-09-05 |
| Context around the text | not documented | checked 2026-09-05 | |
| Telemetry and analytics | None: 'No telemetry, no analytics, error reporting, crash reporting, no usage tracking'; aggregate counters stay on the device | — · — | checked 2026-09-05 |
| Update check | A request with standard HTTP headers, which includes your IP address | Whisperstream's release server, on Cloudflare · automatically; it asks before installing | checked 2026-09-05 |
| Model download | A download request with standard HTTP headers, and no licence key or account identifier | a Cloudflare endpoint · at setup for the speech model, and when you add an optional voice or local AI model | checked 2026-09-05 |
| Account and sign-in | No account. With a Pro licence, the licence key, hostname, OS family and a product identifier go out for validation | Polar, the payment provider · about once every seven days while online with a Pro licence, and when you activate or deactivate a device | checked 2026-09-05 |
| Integrations | not documented | checked 2026-09-05 | |
Facts about Whisperstream checked on 2026-09-05 against whisperstream.io.
Willow Voice (macOS, Windows, iOS)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Your voice, to be turned into text: the help centre says Willow uses cloud-based AI models, and the security page says it runs on cloud infrastructure in the United States. The privacy policy says audio is processed transiently by Willow and its service providers, and the security page that no audio is stored after transcription. A copy stays on your device for re-transcription. | Willow's cloud and its service providers · every dictation, unless you are on the paid plan's offline dictation | checked 2026-09-05 |
| The dictated text | The recognised text is produced in the cloud. In Private Mode, the default, the help centre says no dictated text is collected; if you switch on Help Willow Improve, anonymised recognised text is kept to train its correction models. | Willow's cloud · every dictation | checked 2026-09-05 |
| Context around the text | With Context Awareness switched on, relevant text from the active window, described as small bits of text around your cursor, to improve spelling and formatting; not retained. When it is off, the help centre says Willow does not read any on-screen text. | Willow's cloud · only with Context Awareness on | checked 2026-09-05 |
| Telemetry and analytics | Device type, operating system, app version, performance and diagnostic information, your words-per-minute rate and the number of words dictated | Willow · not stated | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | not documented | checked 2026-09-05 | |
| Account and sign-in | Name, email address, authentication information, account settings and subscription status | Willow · when you create the account and sign in | checked 2026-09-05 |
| Integrations | If you connect a Google account: calendar events, contacts and directory entries, read-only, fetched when needed and discarded after the request | Google, through Willow · only if you connect it | checked 2026-09-05 |
Facts about Willow Voice checked on 2026-09-05 against willowvoice.com.
Windows Voice Typing (Windows 11, Windows 10)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Your voice. Voice typing uses online speech recognition, and Microsoft's privacy page says voice data is sent to Microsoft to provide the service and create the text transcription. Microsoft says it does not store, sample or listen to voice recordings without your permission. | Microsoft's Azure Speech services · every time you dictate; an internet connection is required | checked 2026-09-05 |
| The dictated text | not documented | checked 2026-09-05 | |
| Context around the text | not documented | checked 2026-09-05 | |
| Telemetry and analytics | Nothing unless you opt in to contributing voice clips. If you do, clips are stored de-identified, not associated with your Microsoft account or any other Microsoft ID, kept for up to two years, and may be reviewed by Microsoft employees and vendors to improve speech recognition. | Microsoft · only if you turn on contributing voice clips | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | Nothing to download: voice typing recognises speech online rather than with a model on your PC, which is why it needs an internet connection. | — · — | checked 2026-09-05 |
| Account and sign-in | not documented | checked 2026-09-05 | |
| Integrations | not documented | checked 2026-09-05 | |
Facts about Windows Voice Typing checked on 2026-09-05 against support.microsoft.com.
Wispr Flow (Windows 10 or 11 (x64 only, no ARM), macOS 12 or later, iOS, Android)
| What | What is sent | Where · when | Checked |
|---|---|---|---|
| Audio | Your dictation audio. The privacy page says transcription always happens in the cloud; the security FAQ says the service is hosted with a major US cloud provider and is not end-to-end encrypted, because the audio has to be decrypted to transcribe it. | Wispr's cloud · every dictation | checked 2026-09-05 |
| The dictated text | The transcript comes back from the cloud. With Dictation Cloud Storage on, transcripts, audio, dictation history and personalised speech models are stored on Wispr's servers; with Improve the model for everyone on, audio, transcripts and edits may be used to train its models. | Wispr's servers · every dictation; training is chosen during onboarding, pre-selected, and cloud storage is a setting under Data and Privacy | checked 2026-09-05 |
| Context around the text | With Context Awareness on, the default on Mac and Windows, the help page says context is sent to Wispr with each dictation request: app info, the text in the box before, at and after the cursor, on-screen text, variable and file names in coding apps, your user identifier within the app, the apps in your current session, a screenshot, and conversation history. Password fields, numeric-only fields and browser URL bars are excluded. Windows support is described as partial: conversation context, element descriptions and IDE-specific handling are Mac-only today, and the page does not say which of the other items the Windows app sends. | Wispr · with each dictation request, unless Privacy Mode is on or you turn Context Awareness off under Settings → Data and Privacy | checked 2026-09-05 |
| Telemetry and analytics | The privacy policy says third-party web analytics such as Google Analytics may be used on its Services; it does not itemise what the desktop app itself reports. | Google Analytics · not stated | checked 2026-09-05 |
| Update check | not documented | checked 2026-09-05 | |
| Model download | Nothing to download: transcription always happens in the cloud, so no speech model lives on your PC. | — · — | checked 2026-09-05 |
| Account and sign-in | An account is required: signing in with Google, Apple, Microsoft, SSO, or email and password is a setup step on desktop and mobile. The privacy policy lists email address, first and last name and phone number as account information. | Wispr · at sign-up, before the first dictation | checked 2026-09-05 |
| Integrations | Calendar, contact and directory data from Google, if you connect those; the pricing page lists calendar and Slack connectivity and AI-tool integration on every plan. | Wispr, and the services you connect · only when you connect them | checked 2026-09-05 |
Facts about Wispr Flow checked on 2026-09-05 against wisprflow.ai.
Before you ask
Questions people search for
Does Deixis upload my audio?
No, under any setting. Speech recognition runs inside the app on your own device, whichever of Windows, Mac, iPhone or Android it is, so the recording, the model and the transcript stay there. The same goes for the voiceprints meeting mode keeps to tell speakers apart: they stay on the machine, are never uploaded, and can be deleted per person. There is no code path that sends audio to a network, and no account to attach it to.
Can I see what a tool sends without trusting its privacy policy?
Yes, and it is worth doing. Put a proxy such as mitmproxy between the app and the network, or watch outbound connections with a firewall that logs them, then dictate a sentence. A tool that sends audio produces a large upload on every utterance; a tool that sends screenshots produces periodic image-sized posts. The tables here are what the documentation says; a proxy shows you what the software does.
How often are these tables re-checked?
Every entry carries the date it was checked, and a fact older than about three months raises a warning in our build so it gets re-read. Vendors change their pricing and their privacy terms without announcing it, so the date beside a fact matters more than the fact.
What does Deixis send when cleanup is on?
When you configure a cleanup endpoint and an utterance escalates to it, Deixis sends the transcript, your vocabulary terms, and where the text is going: the focused app's name, the control and field name, and your previous utterance in that field. It never sends the window title, which is the most identifying thing in that structure. With no endpoint set, nothing is sent and the local rules engine does the cleanup.
Why does the table say "not documented" for some rows?
Because the vendor's privacy policy or documentation did not address that row when we read it, on the date shown. We do not guess in either direction. If you know of a page that answers it, send it to support@deixis-voice.com and the row gets updated with the date.
Hold a key. Speak. Keep working.
Deixis is free, needs no account, and runs on your own CPU.
v0.9.0 · 64‑bit Windows 10/11 · 46.7 MB · nothing else to install